Multiresolution wavelet analysis based feature extraction for neural network classification

In this paper we introduce a novel feature extraction scheme as a preprocessor for artificial neural network (ANN) classification. We have shown that the feature extraction scheme implemented via a non-stationary Gaussian Markov random field (GMRF) based on a multiresolution wavelet framework can provide effective features for both the ANN and Fuzzy C-Means (FCM) classification. In our experiment with natural textures and real world digital mammography, each region of the tested images is assumed to be a different class. A label field with each region or class being represented by the same grayscale was then found by the back propagation neural network (BPNN) and FCM clustering algorithm using the extracted discriminatory features. Further enhancement of the segmented result was achieved via Bayesian learning. The formulation of this maximum a posteriori (MAP) estimator was based on the Gibbs prior assumption which is especially appropriate for modeling real world mammograms. Although being estimated by constrained optimization, the MAP estimator can also be found from neural networks such as the Boltzman and the Mean-field-theory machines.